Deep Learning for Motor Brain-Computer Interfaces based on Stereo-Electroencephalography

: (Alternative Format Thesis)

Abstract

Brain-computer interfaces (BCIs) are technologies that have the potential to bypass damaged neural pathways by recording neural signals from the brain and translating them to control external devices. BCIs can be broadly categorized into invasive or non-invasive according to the recording device employed. On the other hand, BCIs also can be grouped by their applications, such as motor, speech, or affective BCIs. This PhD project focuses on motor BCIs based on stereo-electroencephalography (SEEG) signals. In addition, to obtain high decoding performance, this PhD project uses state-of-the-art deep learning methods to decode the SEEG signals.

For motor BCIs, the SEEG paradigm that records the signal from deep regions is less investigated and the decoding performance is still unsatisfactory. To push this line of research, this PhD project focuses on using deep learning methods to enhance the decoding accuracy of the SEEG signals. To accomplish that, various deep-learning models were investigated. It was demonstrated that deep learning was superior in classifying SEEG signals compared to the frequently used ‘traditional’ methods, such as the SVM method. Although a superior result can be obtained using the deep learning method, the data sparsity issue of the SEEG data might deteriorate its performance. To tackle the data sparsity problem, this PhD project investigated different data augmentation methods. It was demonstrated that a transformer-based deep learning model was superior in generating artificial SEEG data and boosting the classification performance of a deep learning classifier based on the convolutional neural network (CNN).

In addition, one particular feature of the SEEG data is the spatial distribution of the recording electrodes which penetrate many deep brain regions, and the encoding of hand/arm movement in the deep brain region is less investigated. Therefore, in this PhD project, we try to use the deep learning method to investigate which brain regions contribute the most to the decoding task. It was demonstrated that most informative electrodes came from the cortical motor-related areas, and only a few electrodes were required to obtain a high performance.

Finally, to further explore the potential of SEEG for motor BCIs, this PhD project investigated the possibility of decoding grasp force from SEEG signals. For the first time, that continuous grasping force can be reconstructed from the SEEG signals.

In conclusion, this PhD project investigated the potential of SEEG-based motor BCIs using deep learning methods. The experiment results showed that SEEG is a promising motor BCIs paradigm and both kinematic and kinetic information can be decoded using deep learning methods. Finally, this PhD project demonstrated that motor decoding can be achieved using only a few electrodes.

Date of Award	24 Apr 2024
Original language	English
Awarding Institution	University of Bath
Supervisor	Dingguo Zhang (Supervisor) & Benjamin Metcalfe (Supervisor)